The run-up to the International Conference on Learning Representations (ICLR) 2026 is showing how quickly generative AI can strain academic peer review. The problem is not simply that researchers are using AI tools. It is that some uses appear to weaken the basic bargain of scholarship: authors should submit work they can defend, and reviewers should evaluate what is actually on the page.
Posts on Reddit and discussions on OpenReview point to frustration on both sides of the process. Authors have been accused of inventing sources, while reviewers have been accused of producing critiques that look as if they were written without reading the work closely.
A trust problem on both sides of peer review
The current review phase has produced examples that go beyond ordinary disagreement between authors and reviewers. In one direction, reviewers have flagged papers that appear to contain AI-generated or otherwise unreliable references. In the other, authors say reviewers are relying on large language models to generate negative reports that miss material already present in the submission.
That combination is damaging because peer review depends on mutual confidence. A reviewer must be able to assume that a bibliography points to real work. An author must be able to assume that criticism is grounded in the submitted paper, not in a generic AI summary or a mistaken prompt response.
The source article describes a community that is not broadly rejecting AI assistance in every form. Using large language models for editing and language polishin, especially for non-native speakers, is generally accepted. The concern is reckless use: allowing AI output to replace verification, reading, and judgment.
Fake citations put one paper under pressure
One case centers on "BrainMIND", a paper from researchers at the Georgia Institute of Technology and China's Tsinghua University. The paper promised an interpretable mapping of brain activity, but reviewers found serious problems in its references.
The reference list included fabricated titles and placeholder names such as "Jane Doe" as co-authors. A reviewer identified the apparent use of a language model and gave the submission a "Strong Reject" recommendation.
The authors then revised the manuscript and its references. But more errors were found afterward, and the authors ultimately withdrew the paper.
The case shows one of the clearest risks of AI-assisted academic writing: a polished-looking manuscript can still contain claims or citations that do not survive basic checking. In a field where papers build on prior work, unreliable references are not a cosmetic flaw. They can make it harder for reviewers and readers to understand what has actually been shown.
Authors push back against AI-written reviews
A separate dispute involved "Efficient Fine‑Tuning of Quantized Models via Adaptive Rank and Bitwidth". The authors withdrew their submission after receiving four rejections, saying the reviewers had used AI tools to generate feedback without reading the paper.
According to the authors, the reviews referred to missing experiments, including GSM8K benchmarks, and unspecified methods that they said were already described in the main text and appendix. In their withdrawal statement, they called the behavior "flagrant desecration of the reviewer's sacred duty" and criticized what they described as AI-induced reviewer laziness.
The complaint highlights a different failure mode. If a reviewer uses AI to speed up the process but does not check whether the resulting critique matches the manuscript, the review can become detached from the evidence. For authors, that can turn peer review into a process that feels arbitrary rather than rigorous.
Even when a paper has weaknesses, a useful review needs to identify them accurately. A mistaken demand for material that is already present wastes time and weakens confidence in the decision.
Publication pressure helps explain the pattern
The source article connects these incidents to a broader study in Research Ethics by Xinqu Zhang and Peng Wang. The study examines how government programs such as China’s "Double First-Class" initiative can produce toxic incentive systems at top universities.
Zhang and Wang describe a process called cengceng jiama, meaning a stepwise intensification of pressure through layers of academic bureaucracy. National policymakers set broad goals such as achieving "world-class status." University leaders translate those goals into ranking targets. Faculty deans then pass them down as stricter publication expectations.
In the study’s account, encouragement for research output can become mandatory SCI journal publication quotas. The researchers call one result "goal-means decoupling": researchers remain focused on meeting output targets while becoming separated from ethical standards.
The study documents cases in which junior scientists said they had "no choice" but to falsify data or hire ghostwriters to keep their positions in a publish-or-perish environment. It also cites data from publisher Hindawi, which in 2023 retracted more than 9,600 papers, about 8,200 of them co-authored by researchers from China.
Why the AI debate is really about accountability
The ICLR 2026 examples show how generative AI can amplify existing weaknesses in academic publishing. If authors are already under pressure to produce more, AI can make it easier to generate text, citations, and revisions quickly. If reviewers are overloaded, AI can make it tempting to produce reports faster than careful reading allows.
But the underlying standard remains the same. Authors are responsible for the accuracy of what they submit. Reviewers are responsible for the substance of the judgments they provide.
The institutional response described in the Research Ethics study makes that harder. To protect rankings and external reputation, university administrators may tolerate unethical behavior as long as results appear favorable. One dean quoted in the study used the Chinese proverb: "Where the water is too clean, there are no fish." The source article summarizes the resulting strategy as turning big problems into small ones and ignoring small problems unless a scandal becomes public.
That is why the current controversy matters beyond any single paper. AI peer review can be useful only when humans remain accountable for the output. Without that accountability, generative AI does not fix the pressure inside research culture. It gives that pressure a faster way to produce errors.